Traceable Claims in Bioinformatics

 Quick Analysis Plan Copied

 Long Analysis Plan

Evidence Base

FibrilNet is a 17,997-node human protein-protein interaction graph (BioGRID, 925,977 interactions) augmented with 9-dimensional Gene Ontology semantic edge features, applied to reconstruct amyloid disease modules via random walk with restart (RWR) FibrilNet primary study.

Reported effect sizes across four modules were substantial and statistically robust, with paired one-sided Wilcoxon tests yielding p values spanning 9.49×10⁻⁴ to 4.46×10⁻²³.

The logarithmic scale is necessary because renal AA semantic-RWR (MRR 0.09061) exceeds the neurologic downregulated baseline by two orders of magnitude. Renal AA showed the largest paired ∆MRR (0.0881, 95% bootstrap CI 0.0282–0.1641), while cardiac ATTR had the highest statistical significance (p = 4.46×10⁻²³, 132/151 targets improved, median semantic/topological rank ratio 0.253).

Computational Pipeline Plan

The reproducible analysis requires five sequential stages, each with explicit validation criteria drawn from the reported data:

• Stage 1 — Graph construction: Parse BioGRID interactome to 17,997 nodes / 925,977 edges; map all node identifiers to UniProt accession via the Ensembl homology service; annotate each edge with a 9-dimensional GO semantic similarity vector derived from Molecular Function, Biological Process, and Cellular Component ontologies.
• Stage 2 — Module definition: Import cardiac ATTR plaque proteome (151 mapped proteins), kidney AA atlas (48 from 418 AA vs 2,130 comparator, FDR<0.05, positive log₂ effect), kidney ALECT2 atlas (124 from 474 vs 1,618), and neurologic ATTRv sural-nerve proteome (202 of 213 humanized via one-to-one Ensembl orthologs; stratify 38 upregulated / 164 downregulated).
• Stage 3 — Diffusion: Implement topology-only RWR and semantic-RWR (transition probability weighted by normalized GO semantic edge similarity) with restart parameter calibrated on the 151-protein cardiac ATTR module.
• Stage 4 — Leave-one-out validation: For each held-out target, rank all 17,997 nodes by RWR score; compute MRR, Recall@100, Recall@200, median rank; apply one-sided paired Wilcoxon signed-rank test with Benjamini-Hochberg correction across modules and Holm correction within neurologic direction strata.
• Stage 5 — Cross-disease overlap: Compute Jaccard similarity and feature-enrichment core (FEC) across the four modules; verify APCS/VTN/TIMP3 at FEC=1 and APOE at FEC=0.75.

Cross-Disease Overlap Structure

The reported Jaccard similarities define a clear tissue-dominant architecture: cardiac ATTR–AL overlap (0.387) is highest because both derive from cardiac tissue, while all cardiac–renal comparisons (0.067–0.136) and cardiac–neurologic comparison (0.0569) are low. This means semantic-RWR gains are concentrated within-tissue, and cross-tissue convergence is restricted to the APCS/VTN/TIMP3 core.

Note: Cardiac–renal pairwise values (0.067–0.136 range) are shown as representative midpoints; cardiac–neurologic = 0.0569 is directly reported. Underspecified pairwise values (AL–AA, AL–ALECT2, AL–neurologic, AA–ALECT2, AA–neurologic, ALECT2–neurologic) are not individually reported and are shown as structural placeholders only.

Known Limitations and Blind Spots

• The four modules derive from different tissues, species (mouse sural nerve for ATTRv), and experimental designs; effect sizes may not transfer across tissue contexts.
• Interaction networks are incomplete and literature-biased; GO annotation depth is non-uniform across protein families.
• The neurologic upregulated subgroup (n=38) showed no significant semantic-RWR gain after Holm correction, indicating the ontology-aware advantage is direction-specific.
• TTR-seeded recovery remains weak even with semantics (MRR 0.005684), confirming that precursor identity is separable from the distributed downstream environment.
• No human variant-resolved ATTRv tissue dataset has been analyzed; the proposed precursor-plus-tissue decomposition remains untested in human neurologic biopsies.

All derived modules, mapping tables, and figure source data are archived at Zenodo (doi:10.5281/zenodo.22161436) for direct reproduction.